基于CNN-MultiLSTM的坍塌压力预测方法  

Collapse Pressure Prediction Method Based on CNN-MultiLSTM

作　　者：张鑫 卢运虎[1,2] 付兴 谢仁军 周长所 袁俊良 宋杨杰 Zhang Xin;Lu Yunhu;Fu Xing;Xie Renjun;Zhou Changsuo;Yuan Junliang;Song Yangjie(State Key Laboratory of Petroleum Resources and Prospecting,China University of Petroleum(Beijing);College of Petroleum Engineering,China University of Petroleum(Beijing);CNOOC Research Institute Co.,Ltd.)

机构地区：[1]中国石油大学(北京)油气资源与探测国家重点实验室,北京市昌平区102249 [2]中国石油大学(北京)石油工程学院 [3]中海油研究总院有限责任公司

出　　处：《石油机械》2025年第2期1-8,共8页China Petroleum Machinery

基　　金：国家自然科学基金面上项目“深层脆性页岩井壁失稳的化学断裂机理与控制”(52074314);国家重点研发技术课题“井壁稳定性闭环响应机制与智能调控方法”(2019YFA0708303-05)。

摘　　要：地层坍塌压力的解析模型参数获取烦琐,加之用于机器学习的随机数据可解释性差,严重影响了钻前预测精度。为此,基于已钻井波阻抗分布规律,生成海量的虚拟井波阻抗数据;利用褶积理论和地层坍塌压力模型计算得到虚拟井的合成地震记录和坍塌压力数据,建立基于CNN-MultiLSTM神经网络的坍塌压力钻前预测模型。研究结果表明:基于已钻井数据生成海量虚拟井训练数据的方法,有效提升了常规随机生成训练数据的可解释性;基于正交试验对神经元个数、学习率、迭代次数等超参数优化后的新模型在训练集和验证集的均方根误差均小于0.1,性能优于随机森林等传统训练模型;对B-8井的坍塌压力(当量密度)的预测值与解析解的绝对误差小于0.02 g/cm^(3),相对误差小于1.635%;井径测井和电成像测井结果显示,预测结果与实钻情形较为一致。研究结果有效提升了利用地震数据预测坍塌压力的预测精度,对推广机器学习方法在石油工程领域的应用具有积极作用。The parameter acquisition for the analytical model of formation collapse pressure is cumbersome,and the interpretability of the random data used for machine learning is poor,which seriously affect the accuracy of predrilling prediction.Based on the wave impedance distribution of existing wells,mass virtual well wave impedance data were generated.Then,the convolution theory and the formation collapse pressure model were used to figure out the synthetic seismogram and collapse pressure data of the virtual well.Finally,a predrilling prediction model of collapse pressure based on CNN-MultiLSTM was built.The research results show that the method of generating mass virtual well training data based on existing well data effectively improves the interpretability of conventional randomly generated training data.The new model built after optimization of hyperparameters such as the number of neurons,learning rate and iteration times based on orthogonal test has a root mean square error of less than 0.1 in both the train and validation sets,and performs better than traditional training models such as random forest. Application to Well B-8 reveals the absolute error and relative error between the predicted collapse pressure and the analytical solution being less than 0.02 g/cm^(3) and less than 1.635%,respectively.The caliper logging and electrical imaging logging results show that the predicted results are relatively consistent the actual drilling findings.The research results effectively improve the accuracy of collapse pressure prediction using seismic data,and play a positive role in promoting the application of machine learning method in the field of petroleum engineering.

关 键 词：坍塌压力 井壁稳定 钻前预测 深度学习 CNN-MultiLSTM神经网络 训练模型 正交试验 

分 类 号：TE242[石油与天然气工程—油气井工程]